Multi-agent LLM middleware: lessons from Obelisk
The middleware chain behind more than 15 production LLM agents: model selection, summarization, tool selection, approvals and retries.
Writing
Notes from the systems we ship: what broke in production and the fix that held, and where a model helps a product and where it gets in the way.
10 posts on 4 topics
Every post
Showing all 10 posts
- Read the post
Muhammad Hassan Raza, Founder
AI agents6 min readPart 1 of 4, Agents in production
LLMs can't save bad UX
Three AI features that failed because the UX underneath was broken, and a five-question test to run before adding an LLM to a product.
Muhammad Hassan Raza, Founder
Product8 min read
Structured logging for AI debugging
Why one structured event per request, with errors that carry a why and a fix, makes logs an AI assistant can actually debug from.
Muhammad Hassan Raza, Founder
AI agents3 min read
Every abstraction I regret
Abstractions I added before I needed them, why each one got worse over time, and the rule I use now before extracting anything.
Muhammad Hassan Raza, Founder
Engineering6 min read
Scoring fraud in legal intake calls
How RISQ scores legal intake calls for fraud using transcripts, external verification, image checks and a chain of disposition gates.
Muhammad Hassan Raza, Founder
AI agents4 min read
I shipped a race condition that double-charged customers, and other war stories
Five production bugs I shipped in Django, PostgreSQL and LangChain code, why each one happened, and the fix that held.
Muhammad Hassan Raza, Founder
ERP and retail7 min read
The AI features users actually want are rarely chatbots
The AI features that keep users are the ones that remove work: smart defaults and background summaries.
Muhammad Hassan Raza, Founder
Product6 min read
Optimizing Django ORM queries for large applications
How I cut query time in a Django POS billing system with select_related, prefetch_related, bulk writes and query profiling.
Muhammad Hassan Raza, Founder
Engineering3 min read
LangGraph vs traditional workflow engines: when to choose agentic AI
A decision framework for choosing between LangGraph's agentic workflows and traditional workflow engines like Airflow or Temporal. Learn when AI agents make sense.
Muhammad Hassan Raza, Founder
AI agents4 min read
Polaris ERP architecture: scaling lessons from Pakistani retail
How Polaris ERP is built: a Vue front end on Django and PostgreSQL, tenant isolation in the ORM, connection pooling, a read replica for reports and staged deploys.
Muhammad Hassan Raza, Founder
ERP and retail7 min read
Agents in production
One agent system followed from its middleware to the logs that make a bad run debuggable.
Running Polaris
The ERP we run for retail businesses: its architecture, its slow queries and the abstractions we would take back.
Questions about any of this go to a person.
There is no mailing list. New posts land on this page and nowhere else. If something here matches a problem you have, book a call and we will tell you whether it applies to your setup or not.
Book a call- First reply
- same working day
- Hours
- Set to your time zone
- Built in
- Lahore, Pakistan